The Second Scaling Law of AI remains undefeated. If you want better hacking (or math, or science, or crossword puzzle solving) out of an LLM, just let it use more tokens. There doesn't seem to be any plateau so far in the new study by the UK's governmental AI Security Institute.

I'm never selling my 2017 Taco

This is going to be a very long shot give my audience, but... Any good alternative to Caltopo? I don't have any of the accounts they require for login, cannot use email for whatever reason, and I don't want to create an account just for this

“Training language models to be warm can reduce accuracy and increase sycophancy” Who would have thought?

Read this yesterday while it was trending on HN As someone who moved his entire cloud stack to EU only providers: yeah, I fully agree I’m never going back

Sub cancelled. Feedback left. This isn't going to change anything I guess, but it's worth a shot...

So now I have a date by which I will switch provider cause Claude won't be usable for me... ☹️ At least they gave a heads up I guess

I need more time to read all the papers that are constantly coming out lol

I did a (potentially useful to some) thing! Multilingual tokenizers break Latin into 3-5 subwords per word because Latin is like <0.01% of the training data. So I trained one on 738M words of actual Latin. ~1 token per word huggingface.co/v37/latin-to...

“cuda-oxide is an experimental Rust-to-CUDA compiler that lets you write (SIMT) GPU kernels in safe(ish), idiomatic Rust. It compiles standard Rust code directly to PTX — no DSLs, no foreign language bindings, just Rust.”

Today’s project was building a coop for the pullets. Normal programming resumes on Monday.

The cool superpower you get with LLMs is that you can have them find relevant research papers, explain them in plain Italian/English, and help you implement them. Lares happened to find a papers on its own after our attempt at SFT didn't work for voice transfer. Linking below.

While I work, Lares is off doing its own thing: 1. Vector steering Qwen 3.6 MoE based on our chats from the last few months 2. Training a 300M model on classical Latin only 3. Filter/dedup 95GB of Latin texts to train a tokenizer, then a bigger Latin model The beauty of 256GB RAM + GPUs to spare

So this just happened

This is really cool Models trained specifically to predict a larger model’s output as opposed to the classic speculative decoding setup that uses a normal smaller model And they even the KV cache

Finally got to try pi.dev and I'm pretty impressed by how customizable it is Which means if you know how to optimize your harness you can get LLMs to actually function better than on their native harnesses

I usually don't post political stuff but this is hilarious 😂 Anyways, back to ignoring them and living a happy life

Yep, this is the way

I’m getting more and more frustrated with Opus 4.7, and I wish I could go back to 4.6 Is it just me? It might be more accurate, but it’s so annoying to use

I see you and I respect you

"LLMs can hide text in other text of the same length" And 8B open weights models are apparently good enough arxiv.org/pdf/2510.20075

Second run of QLoRA on Qwen 3.6 using 4 months of convos with @laresai.danielesalatti.it (plus its notes and letters to itself) as training material Gotta say this stuff is fascinating I have no business doing any of this but I can cause the little genie in the CLI is teaching me

The more software you can make, the more software you need? www.citadelsecurities.com/news-and-ins...

new post answering @aaronsterling.bsky.social's tag on why companies keep going all-in on AI even when the ROI isn't there. my POV: the tech works, the gap is everywhere else. uses an org model to diagnose, with a nod to @timkellogg.me from the eng side. www.appliedaiformops.com/p/the-gap-be...

It looks like it was a bug in some abuse classifier and are refunding users Ok, that's good to see

Did not know SIMD was a thing, this is cool lemire.me/blog/2026/04...

This is starting to happen a little too frequently semgrep.dev/blog/2026/ma...

I really hope that: 1. this is a bug 2. they will promptly rectify and refund the users affected Because if it's not, this behavior from a company is reason enough to move away (regardless of wether you use OpenClaw and are affected or not) news.ycombinator.com/item?id=4796...

I think I’m going to get my hands on Mistral/DeepSeek/Qwen/Gemma and see how they do here: github.com/lsdefine/Gen...

Mistral Medium 3.5 is out

RIP GitHub I’m in the process of moving everything public to Coderberg Other stuff will be in a privately hosted forge I don’t see myself going back to GitHub at this point

How do I get my company to expense that for me?

Was going to write a joke that Ethernet is basically Wi-Fi by this definition Looked it up, apparently it was named after the "luminiferous ether" in 1973 on purpose So, yeah...

This is pure evil. You should definitely not visit sci-bot.ru

Hot take (probably not): Claude Code sessions should automatically be named from the tmux sessions they are in

Anthropic is really doing all it can to make me switch to OpenAI

Wait, people use Meta AI?

The usual conversation we should have as a society that instead we are not having

I’m at the point where I have so many things I want to build that Claude Code is too slow Even with 4 to 6 terminals open constantly switching and prompting

DeepSeek v4 just dropped So many new models to play with And so many are open weights!

Qwen 3.5 dense -> Qwen 3.6 MoE -> Qwen 3.6 dense In like 3 weeks? The MoE makes sense, but the two dense models back to back? 🤔

Interesting Token classification afaik is usually the domain of small dense models (DistilBERT, DeBERTa-v3). OpenAI went with 128 experts, top-4 routing, 50M active out of 1.5B total 🤔 I need to read more

Been really eyeing @frame.work laptops to replace my Mac and @fairphone.com to replace my iPhone Not sure why but Apple's walled garden has just felt frustratingly worse lately

A new Git version just dropped and it comes with a new experimental `history` command! `reword` can be used to change commit messages and `split` can untangle a single commit into multiple ones. No more interactive rebase. 🎉 github.blog/open-source/...

Current power consumption during the first LoRA attempt

Been running Lares locally on 2xMI100s Works great I haven’t needed to turn on the heat in my office for the last week IYKYK

Promo Driven Development → Token Count Driven Development. Obviously. You optimize for what you're evaluated on.

Is there a long form blogging AT proto thing somewhere that is self hosted? I want to be able to fully tweak it (theme, style, random pages with custom code that pulls from APIs, etc.)

This eerily resembles asbestos in the way it’s described @esa.int www.esa.int/Science_Expl...